CAN PRIMARY CARE ELECTRONIC MEDICAL RECORDS PREDICT PERSISTENT FREQUENT ATTENDERS? APPLICATION OF A TABULAR FOUNDATION MODEL (TABPFN) TO THIN® PRIMARY CARE DB IN BELGIUM AND SPAIN
Author(s)
Caroline Eteve-Pitsaer1, Elena Zanzottera Ferrari, MSc2, Carlos Iglesias, MD3, Samuel Brouyere, MSc4.
1European RWD-E Analytics Director, Cegedim Health data - Clinityx by Gers Data, Boulogne-Billancourt, France, 2GERS/Cegedim Health Data, Milano, Italy, 3CEGEDIM HEALTH DATA, Sant Cugat del Vallès, Spain, 4GERS/Cegedim Health Data, Anderlecht, Belgium.
1European RWD-E Analytics Director, Cegedim Health data - Clinityx by Gers Data, Boulogne-Billancourt, France, 2GERS/Cegedim Health Data, Milano, Italy, 3CEGEDIM HEALTH DATA, Sant Cugat del Vallès, Spain, 4GERS/Cegedim Health Data, Anderlecht, Belgium.
OBJECTIVES: Frequent attenders (FAs) in primary care drive a disproportionate share of consultations. This study evaluates the individual-level predictive value of routinely collected Electronic Medical Records (EMRs) from the THIN® database to forecast future persistent frequent attendance (pFA) using TabPFN, a novel tabular foundation model.
METHODS: A retrospective adult cohort (N= 923,708) from Belgium and Spain was analysed. The binary outcome, pFA, was defined as falling within the top decile of primary care attendance in each of three consecutive years. Features derived from a 5-year lookback included demographics, chronic comorbidities, prescription counts, sick leave, laboratory tests, and prior physician/nurse visits. TabPFN, a transformer pre-trained on synthetic tasks that predicts via in-context learning without dataset-specific tuning or imputation, was validated using stratified 5-fold and leave-one-country-out cross-validation against logistic regression and HistGradientBoosting.
RESULTS: The pooled prevalence of pFA was 3.24%. Compared to non-FAs, pFAs were significantly older (mean 65.9 ± 16.8 vs 50,5 ± 17,5 years, p< 0.001) and predominantly female (62.4%, p< 0.001). Marked differences were observed in lookback physician consultations (mean 122.0 ± 72.7 vs 26.9 ± 27.0, p< 0.001), nurse visits (67.0 ± 75.4 vs 14.3 ± 18.4, p< 0.001), and endocrine/metabolic comorbidities (79.3% vs 33.9%, p< 0.001). TabPFN achieved an AUROC of 0.956 ± 0.003
and an AUPRC of 0.623 ± 0.012, outperforming logistic regression (AUROC: 0.947) and HistGradientBoosting (AUROC: 0,948), whilst minimising calibration error (Brier score: 0,033). At a 70% target recall operating point, TabPFN flagged only 4.6% of the total population, achieving a Positive Predictive Value of 50.3% and a specificity of 97.6%. Prior physician and nurse encounters were the strongest predictors, and cross-country transportability was excellent (held-out AUROC: Belgium =0.945; Spain =0.950).
CONCLUSIONS: Routine EMRs enable accurate, transferable TabPFN-driven risk-stratification, transitioning health systems toward proactive resource allocation.
METHODS: A retrospective adult cohort (N= 923,708) from Belgium and Spain was analysed. The binary outcome, pFA, was defined as falling within the top decile of primary care attendance in each of three consecutive years. Features derived from a 5-year lookback included demographics, chronic comorbidities, prescription counts, sick leave, laboratory tests, and prior physician/nurse visits. TabPFN, a transformer pre-trained on synthetic tasks that predicts via in-context learning without dataset-specific tuning or imputation, was validated using stratified 5-fold and leave-one-country-out cross-validation against logistic regression and HistGradientBoosting.
RESULTS: The pooled prevalence of pFA was 3.24%. Compared to non-FAs, pFAs were significantly older (mean 65.9 ± 16.8 vs 50,5 ± 17,5 years, p< 0.001) and predominantly female (62.4%, p< 0.001). Marked differences were observed in lookback physician consultations (mean 122.0 ± 72.7 vs 26.9 ± 27.0, p< 0.001), nurse visits (67.0 ± 75.4 vs 14.3 ± 18.4, p< 0.001), and endocrine/metabolic comorbidities (79.3% vs 33.9%, p< 0.001). TabPFN achieved an AUROC of 0.956 ± 0.003
and an AUPRC of 0.623 ± 0.012, outperforming logistic regression (AUROC: 0.947) and HistGradientBoosting (AUROC: 0,948), whilst minimising calibration error (Brier score: 0,033). At a 70% target recall operating point, TabPFN flagged only 4.6% of the total population, achieving a Positive Predictive Value of 50.3% and a specificity of 97.6%. Prior physician and nurse encounters were the strongest predictors, and cross-country transportability was excellent (held-out AUROC: Belgium =0.945; Spain =0.950).CONCLUSIONS: Routine EMRs enable accurate, transferable TabPFN-driven risk-stratification, transitioning health systems toward proactive resource allocation.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR161
Topic
Methodological & Statistical Research
Disease
No Additional Disease & Conditions/Specialized Treatment Areas